Re: pytorch and CUDA
On Wed, 2023-03-22 at 13:17 +0200, Andrius Merkys wrote:
> Hello,
>
> I saw your message on debian-mentors@. Maybe a quick fix would be to
> skip dh_shlibdeps step entirely for now.
The solution is to build a fake library and overwrite it with the downloader.
The solution has already landed onto experimental as of
nvidia-cudnn (8.7.0.84~cuda11.8) experimental; urgency=medium
> > After sorting everything out, I plan to gradually get the remaining
> > reverse dependencies for the cuda variant into experimental.
> > The code on the pytorch-cuda branch will be merged into master.
> > I'll create a new `cuda` branch for the real packaging that will be
> > uploaded to the archive (I need to double check the licenses before
> > doing this).
>
>
> I recently started getting FTBFS for tensorpipe due to nonexistent
> cuda-gcc and cuda-g++. I see they come from nvidia-cuda-toolkit-gcc
> which is commented out in [1]. So 'cuda' branch no longer builds the
> cuda version, is this intentional?
This is neither a issue, nor intentional. I was just working on this
yesterday, and I plan to upload to cuda version of tensorpipe
to experimental.
Before the upload, I try to merge all things from the cuda branch
as comments in the main branch, thus the difference between
the original version and the cuda version can be minimized.
So you are just trying to build amid my finalization of the cuda
variant :-) FTBFS during this stage are easy to fix. Or you can wait
untill you see my NEW queue upload.
Reply to: